NVIDIA added nemotron-3.5-embed-1b
What to do: Call it as nemotron-3.5-embed-1b.
AIWire: AI model shutdowns, prices and API changes
The wire
Every change we detect, newest first. Times are Eastern (ET). Archive
TuesdayOctober 20, 2026
Anthropic Python SDK 1.17.0
What to do: Update to anthropic-python 1.17.0 if you use the Admin API webhooks/RBAC features or Bedrock, Mantle, or Google Cloud clients.
You can now choose 12-hour prompt caching in the API.
What to do: Consider enabling 12-hour prompt caching for workloads with infrequent repeated requests, and factor the 2.5x cache-write cost into pricing.
Anthropic Python SDK 1.16.0
What to do: Update to SDK v1.16.0 to pick up the new Admin API, analytics, and Google Cloud endpoint features.
Inference hooks are out of beta for Claude Enterprise organizations.
What to do: If you're on a Claude Enterprise plan, you can start using inference hooks in production.
Mistral added voxtral-mini-tts-3-0
What to do: Call it as voxtral-mini-tts-3-0.
Mistral added voxtral-mini-transcribe-realtime-3-0
What to do: Call it as voxtral-mini-transcribe-realtime-3-0.
Mistral added voxtral-small-transcribe-3-0
What to do: Call it as voxtral-small-transcribe-3-0.
Mistral added voxtral-mini-transcribe-3-0
What to do: Call it as voxtral-mini-transcribe-3-0.
DeepSeek just open-sourced its inference engine
InfoReleasesHacker News · community
What to do: If you want to self-host DeepSeek models, try DeepInfer from source (Python 3.12, CUDA 13 or CANN 9.2) and plan to supply your own KV cache and load balancing.
The Claude Tag API is available in beta for Claude Enterprise organizations with the experimental-claude-tag-2026-10-21 beta header.
What to do: If you're on Claude Enterprise and manage Slack Claude Tag settings, request a scoped Admin API key and try the beta endpoints with the experimental-claude-tag-2026-10-21 header.
MondayOctober 19, 2026
Step by Step guide: Build a Voice Agent That Never Goes Silent using ADK
What to do: If you build voice agents on the Gemini Live API with ADK, adopt streaming tools or response_scheduling to avoid dead air during slow tool calls.
Launch HN: Laminar (YC S24) – RL-trained model that finds errors in agent traces
InfoReleasesHacker News · community
What to do: Developers building agents who want full-trace error analysis without sampling can try flow-1 via Laminar or self-host the open-source platform.
Maxwell: A new kind of model that sits between code and hardware
InfoReleasesHacker News · community
What to do: If interested, request early access via the Maxwell site, or try the open-source optimised LTX-2.5 pipeline on your own workloads.
nemotron-3-super-120b-a12b has been retired
What to do: Requests to nemotron-3-super-120b-a12b now fail. Switch to another model.
SundayOctober 18, 2026
DeepSeek search and fetch – $0.75 per 1000 successful requests
InfoPricingHacker News · community
What to do: If you use Search or Fetch via DeepSeek Harness, Claude Code, or Codex, plan for per-request billing starting 00:00 Beijing time on November 1, 2026, and review the account-level QPS and Fetch per-do…
SaturdayOctober 17, 2026
OpenAI Python SDK 3.35.0
What to do: Update to openai-python 3.35.0 if you use Live/realtime features or want the ultrafast Agents service tier.
OpenAI Python SDK 3.34.0
What to do: Update the OpenAI Python SDK to 3.34.0 if you need the new Realtime Translation WebSocket sessions or external storage status fields.
Opening up the in-house search API; DeepSeek-V4.1-Flash-1017 Update
What to do: Upgrade to DeepSeek Harness 0.2.1-rc.1 or later for the best experience with the new model.
FridayOctober 16, 2026
OpenAI Python SDK 3.33.0
What to do: Update to openai-python 3.33.0 if you need the new image safety identifier fields or Decisions cache breakpoint typing.
OpenAI Python SDK 3.32.0
What to do: Update to SDK 3.32.0 to access the new cache, residency, voice, and Live session features.
Released prompt-to-voice for all API customers.
What to do: Try the prompt-to-voice endpoints to create custom voices, noting that speech generation is billed under regular GPT-Live session pricing.
Show HN: Gemini 4 Argon plays StarCraft
InfoResearchHacker News · community
What to do: No action needed; background reading.
Google raises gemini-4-argon prices from Dec 1, 2026: cached input $0.10 → $0.20 per 1M tokens; batch, flex and priority tiers too
What to do: Check your cost estimates before Dec 1, 2026.
OpenAI-hosted sandboxes in the Agents API now include the Artifact SDK and skills for creating and editing spreadsheets, presentations, and Word documents.
What to do: Explore the new Artifact SDK and skills in the Agents API sandboxes, and consider setting agent.service_tier to "ultrafast" for lower token latency.
ThursdayOctober 15, 2026
Added native Decisions integration in GPT-Live.
What to do: Explore wiring Decisions into your Live sessions via native delegation or transcript delta triggers.
Fast mode for Claude Opus 5.5 is now available on Claude in Microsoft Foundry for deployments hosted on Azure.
What to do: If you use fast mode with Claude Opus 5.5 on Azure-hosted Microsoft Foundry deployments, you can drop the beta header, though requests that still send it keep working.
Fast mode is out of research preview for Claude Opus 5.5, and you no longer need to request access.
What to do: Remove the fast-mode beta header from Claude Opus 5.5 requests if you no longer need it, but keep sending it for Claude Opus 5 and Claude Opus 4.8 fast mode requests.